yes I read that. do you think it's reasonable to assume that the same expert will be selected so consistently that model swapping times won't dominate total runtime?